📦 slee-persis/gvs5h

GVS5H: Five Qwen Models Match Claude Fable 5

GVS5H is the research project that proves five smaller models working together can match top-tier coding performance. Instead of relying on one massive, expensive model, this method splits a hard coding problem into smaller pieces, letting a group of models share notes and plans through a simple file system. This coordinated approach allows standard open-weight models to solve complex coding challenges with accuracy that rivals the most expensive proprietary tools. It is a fascinating look at how organization and teamwork can boost AI performance without needing massive hardware or huge budgets.

🆔 @hackernewsgithubprojects

📦 tonyd2wild/deepseek-v4.1-flash-vllm-dgx-spark

552B Model on Four Spark Devices

A five-hundred-and-fifty-two-billion parameter AI model now runs on just four compact desktop computers. The DeepSeek V4.1 Flash project takes a massive language model that normally requires a data center and fits it onto a small cluster of NVIDIA DGX Spark boxes. The core challenge was memory. The model has a huge internal dictionary that does not fit in standard memory, so the developers kept that dictionary on a hard drive and only pulled specific lines when needed. This required fixing several software bugs to get the hardware talking correctly.

🆔 @hackernewsgithubprojects

📦 ship-studio/ship-studio

Ship Studio: The AI Developer Studio

Ship Studio is the local first development app that puts your favorite command line coding agents right inside your project window. Instead of juggling separate terminals, browsers, and hosting dashboards, you get one clean workspace where you can talk to tools like Claude Code or Codex directly next to your live preview. It’s free, runs entirely on your machine, and handles the boring stuff like git branches and deploying your site automatically. You don’t have to learn a new syntax or leave your comfort zone to make changes. Just describe what you want, watch the agent build it, and push it to the world.

🆔 @hackernewsgithubprojects

📦 xuzhougeng/scientificfigurelibrary

Scientific Figure Library: Your Plotting Vault

Your scientific figures live on your hard drive, not in the cloud. Scientific Figure Library acts as a personal vault where you import plots and their code, review them, and publish exact, immutable versions to a single local folder. This solves the chaos of hunting through old scripts for that perfect graph layout. You simply search your library, confirm a template, and drop the exact code and image into any new project. It works directly with Claude Science and Wisp Science, so your coding agents can reuse your approved figures without guessing. It never runs the code for you, keeping your data safe and your workflow predictable.

🆔 @hackernewsgithubprojects

📦 narumiruna/pi-extensions

Pi Extensions: Make Your Coding Agent Your Own

Customize your coding agent's personality and workflow by picking and choosing exactly which tools it needs. Pi Extensions is a collection of add-ons that turns a basic AI assistant into a highly personal productivity machine. The coolest part is the goal mode, which forces the agent to keep working on a task until it is truly finished, rather than stopping halfway. It is like having a persistent intern who refuses to go home until the job is done. This repo solves the problem of generic AI tools that do not fit your specific coding style.

🆔 @hackernewsgithubprojects

📦 ion-elgreco/rivers

Rivers: Rust-Powered Python Pipelines

Rivers is the orchestration platform that keeps your data pipelines blazingly fast without sacrificing Python simplicity. Most tools run the heavy logic in Python, which can get slow, but Rivers uses Rust for the core engine. You still write your pipelines in Python, but the tricky parts like figuring out what runs next happen in compiled code. This means the system can handle massive, complex workflows without breaking a sweat. If you are building data systems that need serious speed and reliability, this is a tool worth checking out.

📰 https://news.ycombinator.com/item?id=49044862

🆔 @hackernewsgithubprojects

📦 taruma/sceneflow

SceneFlow: The Script-to-Screen Sync Tool

SceneFlow is the script-to-screen sync tool that finally lets you see exactly which part of your prompt is on screen at any given moment. Built for AI filmmakers, it links specific lines of your screenplay to video timestamps so you can watch your script highlight in real-time as the footage plays. It handles both traditional screenplays and technical auteur scripts, helping you spot missed instructions or camera drift instantly. This makes it incredibly useful for checking if your AI video model actually followed your creative vision without manual guesswork. If you are generating video from text, this is the quickest way to know what worked and what needs fixing.

🆔 @hackernewsgithubprojects

📦 remvisual/claude-handoff

Claude Handoff: Never Lose Your Code Context

Save your coding progress before the AI forgets it. Claude Handoff captures every decision, failed attempt, and measurement from your session into a single, structured note. This solves the frustrating problem of starting each new AI coding session from scratch, wasting time rediscovering what was already tried. The tool automatically links these notes together, so your third session knows exactly what happened in the first two. It turns a chaotic conversation into a clear roadmap, ensuring your next session picks up precisely where you left off without confusion. Try it once, and you’ll never start a coding session blind again.

🆔 @hackernewsgithubprojects

📦 see-sol-lab/deepseekgui

DeepSeekGUI: The AI Workbench That Actually Works

DeepSeekGUI puts a full, self-contained AI coding workspace directly onto your desktop, solving the headache of juggling separate terminals and browsers. Built on the stable DeepSeek Harness, this tool lets you chat with an assistant that reads your code, edits files, and runs commands while keeping everything local for privacy. The coolest part is the integrated Git and browser view, which lets you see exactly what the AI changed and browse web pages right next to your conversation. It even remembers your project details across sessions, so you never have to repeat yourself.

📰 https://news.ycombinator.com/item?id=49488865

🆔 @hackernewsgithubprojects

📦 featrs/featrs

featrs: The Rust Data Prep Powerhouse

featrs brings the massive, familiar preprocessing toolkit of Python to the high-performance world of Rust. Built entirely on top of the Polars data engine, it lets you scale, encode, and clean your data frames without ever leaving the Rust ecosystem. The real surprise here is that it doesn't just do the math; it handles the messy, real-world headaches like missing indicators, rare category grouping, and even complex time-series lags and rolling windows. It mirrors the structure of scikit-learn so you can translate your mental models directly into Rust code. This is the missing piece for data engineers who want the speed of Rust and the structure of Python.

🆔 @hackernewsgithubprojects

📦 xnet-ngo/aiope

One Founder, One AI, One App

Take a full Linux terminal, real-time voice, and a browser, and put them all in the pocket of an Android phone. That is exactly what AIOPE does. It is not just a chatbot that answers questions; it is an autonomous agent that actually gets things done. Ask it to text a friend, check the weather, or write a piece of code, and it will do the work right there on your device. The coolest part is how much power it packs into a single app. It runs a real operating system, automates your web browser, and connects to remote servers, all while keeping your data private.

📰 https://news.ycombinator.com/item?id=49667587

🆔 @hackernewsgithubprojects

📦 miaai-lab/deepseek-v4.1-flash-dgx-sparks

Run DeepSeek V4.1 Flash on Three DGX Sparks

Run a massive language model on just three affordable NVIDIA DGX Spark boxes. This project connects three separate computers into a single, powerful server that handles complex tasks like vision and tool use. The trick is moving huge memory tables to the hard drive so the actual AI fits inside the limited memory of these small boxes. It works by splitting the model across the machines and using a fast network link to keep them in sync. You get a standard API endpoint that is ready to use, which is huge for anyone wanting to test a big model without buying a million-dollar server.

🆔 @hackernewsgithubprojects

📦 vesaias/jobnavigator

JobNavigator: The AI That Hunts Jobs For You

JobNavigator is a self-hosted tool that turns your job hunt into an automated assembly line, running entirely on your local hardware. It scrapes career pages and job boards, then uses AI to score every listing against your résumés to see if you are actually a fit. The coolest part is the Chrome extension: it auto-fills application forms and answers tricky interview questions using your personal profile, so you never stare at a blank text box again. It even tracks your status and pings you via Telegram when a reply arrives. You are in control of the data, the pace, and the privacy.

📰 https://news.ycombinator.com/item?id=49615385

🆔 @hackernewsgithubprojects

📦 crwdla/tokentab

How much do AI tools actually cost?

Tokentab is the offline cost tracker that finally shows you exactly what your AI coding assistants are spending. It reads the session logs already saved on your computer for tools like Claude Code and Gemini, then adds up the token usage to calculate your real costs by model, project, and day. Because it runs entirely locally without needing an API key or internet connection, your data never leaves your machine. You can view these numbers in a simple terminal or a local web dashboard to see exactly where your money goes. It is the straightforward way to stop guessing and start seeing the true price of your automation.

🆔 @hackernewsgithubprojects

📦 0xbakeer/deepseek-v41-flash-spark

DeepSeek-V4.1-Flash Runs on One Box

DeepSeek-V4.1-Flash is the single-machine serving recipe that keeps full quality without shrinking the model. Most attempts to run this massive language model on one computer force you to lower the accuracy so it fits in memory, but this project takes a different path. It holds the most important parts of the model in fast memory while streaming the rest from the hard drive as needed. This means you get the exact same output quality as the original, even if it moves slower. The repository handles all the complex setup so you can connect your favorite chat apps and start using it immediately.

🆔 @hackernewsgithubprojects

📦 kennethwolters/litelm

Litelm: The Lean LLM Router

Route AI model calls across nineteen different providers using a single, unified function call. The popular LiteLLM library handles this job, but it carries a massive overhead of extra features like proxy servers, complex caching layers, and detailed cost tracking that most developers never actually need. Litelm solves this by stripping away all that unnecessary weight. It keeps only the essential routing logic and message translation required to talk to various AI models. The entire project fits in under three thousand lines of code with just two dependencies. If you want a fast, lightweight way to swap between different AI providers without the heavy baggage, this is your streamlined alternative.

📰 https://news.ycombinator.com/item?id=49662767

🆔 @hackernewsgithubprojects

📦 i-am-manware/manware-s-ai-learning-toolkit

Manware's AI Learning Toolkit: The Tutor, Not the Coder

Manware's AI Learning Toolkit is the AI setup that makes your coding assistant act like a stubborn teacher instead of a code generator. Usually, AI tools hand you the answer, but this project flips that completely by forcing you to try first. It uses simple text prompts to make the AI ask you questions and give tiny hints, one level at a time, until you figure out the fix yourself. This keeps you in the driver's seat so you actually learn why the code works, not just how to paste it. It is a clever way to use AI for genuine skill building rather than just getting work done faster.

🆔 @hackernewsgithubprojects

📦 axiomatic-ai/axqm

AxQM: The Quantum Proof Benchmark

AxQM is the quantum benchmark that turns textbook exercises into formal code. It takes one thousand problems from the famous quantum computing book and rewrites them as logic puzzles inside a proof assistant called Lean. Each puzzle starts with a claim and an empty space where the proof should go. Your job is to fill that gap using only verified math, with no shortcuts allowed. The project tracks how hard each puzzle is and how they depend on one another. This gives researchers a clean way to test if artificial intelligence can actually reason through complex physics. It is not just a test set.

🆔 @hackernewsgithubprojects

📦 stretchcloud/campfire

Campfire: The Team Chat for AI Coders

Campfire is the collaborative workspace that turns separate AI coding tools into a shared team brain. Instead of juggling different apps for Claude or Codex, you get one browser window where your whole team can watch agents work in real time. The coolest part is the collective intelligence layer, where the system remembers what the agents learn and shares those insights across your entire project automatically. This means you are not just running bots, you are building a shared memory that gets smarter every time your team codes.

🆔 @hackernewsgithubprojects

📦 bzdmathclub/bzd-math-modeling-skills

The Judge in Your Pocket: bzd-math-modeling-skills

bzd-math-modeling-skills is the AI reviewer that finally replicates the exact judgment of a national competition judge. It analyzes your paper using the specific grading criteria from the last five years, accounting for the subjective factors that usually make scoring difficult. This tool is useful because it gives you an honest position estimate, so you know exactly where you stand before the official results are announced. It is not just a spell checker, but a complete simulation of the review process. If you are preparing for a competition, this gives you a clear and realistic picture of your work's quality.

🆔 @hackernewsgithubprojects